What was done #
A computational study uses language-model neurons and comparative phylogenetic methods to trace recurring story elements across 32 Batman works, then tests whether the signal exceeds ordinary word overlap.
This research project treats a neuron activation in a language model as a candidate unit of cultural inheritance and asks whether those units can trace recurring ideas across Batman stories. The analysis covers 32 works, from the 1943 serial to The Batman (2022), and adapts comparative phylogenetic methods previously used to study folktales. The repository includes the paper, an interactive report, and a viewer for the Claude Code conversation used to design, run, check and write up the project.
Findings reported by the authors
The paper reports recurring neuron-defined families across the Batman corpus, including a smaller core appearing in many works. It compares those results with random pseudo-franchises and with plain word-overlap baselines. The authors find that much of the apparent signal is also captured by word similarity: Qwen’s neuron results do not add a reliable residual signal after that control, while Llama shows a narrower result beyond shared words. The write-up also reports robustness checks across analysis specifications and an independent rebuild.
These caveats matter: the project does not claim that every neuron corresponds to a stable cultural idea. Its results depend on model choice, thresholds and how a unit is labeled. The repo provides the paper and analysis materials so readers can inspect those decisions.
AI contribution
The authors publish the Claude Code conversation and describe Claude and Mike working together to design, run, check and write up the project. The experiments also inspect activations from Qwen3 and Llama language models. This is a research entry about model representations, rather than reverse engineering a conventional binary or device.